CtrlK
BlogDocsLog inGet started
Tessl Logo

ainativedev/latest-aidevcon-speakers-london-2026

AI Native DevCon 2026 London — all conference sessions as interactive skills

71

Quality

89%

Does it follow best practices?

Impact

No eval scenarios have been run

SecuritybySnyk

Risky

Do not use without reviewing

Overview
Quality
Evals
Security
Files

outline.mdtalk-batey-building-product-teams-age-of-ai/

Outline — Building Product Teams in the Age of AI

Speaker

Christopher Batey — CTO, Core Engineering Consulting Group. Over 20 years across software delivery, product development, JVM library development (concurrency/distributed systems), infrastructure engineering, architecture & system design, and most recently platform engineering. Open-source contributions including Akka and Kubernetes. Focus: speed and stability — what to do next to speed up the software delivery lifecycle.

Hosted by Katie Roberts on the Latent Space stage (the transcript opens with her introduction; she also says "Thanks very much, Dave" — likely a transcription artifact, since the speaker is Christopher).

Abstract (as provided by user)

Over the last year, I've built and scaled a new product engineering team from scratch: hiring, training, and shipping a real product while AI-assisted development rapidly changed how software gets built. ... This talk shares three practical pillars we've learned for running product development in the age of AI:

  1. Path to Production at AI Speed
  2. Training and Evaluating AI-Enabled Engineers
  3. Designing Workflow for Parallel Change

This isn't a theoretical talk about AI replacing engineers. It's a practical account of what changed when a real team tried to ship real software while development speed fundamentally increased.

Thesis (synthesised)

AI is an amplifier, not a replacement: it dramatically accelerates code production but exposes — and worsens — every weakness elsewhere in the delivery lifecycle (product decisions, adoption, shared system understanding, review capacity). The fix isn't more tooling; it's disciplined practice: refuse to delegate systems thinking to agents, write human-authored ADRs before implementation, keep teams small with adoption-bound missions, work in parallel but limit each engineer to one complex cognitive task at a time, and measure adoption rather than vanity metrics like commits or PRs merged.

Section TOC

§HeadingSummaryTranscript lines
1Intro & company contextHost hand-off; Core Engineering Consulting Group's history as a platform/CD consultancy; how internal training/processes/reference impls looked two years ago.~1–35
2Where they are now (with AI)Interactive training platform that provisions real environments; maturity tooling became SaaS; ref impls became a product for SMEs; the consultancy is its own best customer.~35–65
3The stakes — sensitive systemsTwo product types: one holding sensitive client data, one managing client production environments. The talk is scoped to systems where it really matters if things go wrong.~65–80
4The fear: software no human has seenCode can do anything with data it has access to; banks/healthcare doing this is scary.~80–95
5Audience pollsWhat % of your code is written by AI? (most: 50–100%) and what % goes to production with no human review? Batey's claim: at his company, 90–100% is AI-written, but ~0% goes without human critical thinking.~95–125
6AI as amplifierAI amplifies what you already do well and what you do badly. Enables delivery-focused engineers. The deeper worry: what % of your system goes to production without human understanding?~125–145
7Don't delegate systems thinking to agentsConcrete example: managed service + client-side deployment + auth boundary. Lists system-level questions a 7,000-line PR needs answered.~145–195
8ADR-first workflowAgents wrote verbose ADRs no one read. Fix: humans write the ADR first with diagrams, review it carefully; then agents are good at reviewing follow-up PRs against it.~195–220
9The producer "black box" — harness, host, modelThree layers of opacity. Choose open-source harnesses / alternative hosts / open-weight models to switch providers quickly during outages. Vendor lock-in is fine if it's a conscious choice; baking it into your product is a different game.~220–270
10The whole flow: idea → product mgmt → ownership → build → adoptionDifferent stages get different AI speed-ups. Building got an order-of-magnitude faster; product decisions did not. Look for friction caused by uneven acceleration.~270–305
11"You build it, you run it, you drive adoption"Engineers must drive adoption of the feature they built before starting the next one. Otherwise they build features for fun.~305–325
12Engineer day-to-day: plan / execute / refinePlan and refine use the human brain; execute is what AI accelerates. Refinement is the magic — fast user-feedback iteration. Superpowers and spec-driven frameworks help.~325–355
13Parallelism — but one complex task at a timeEveryone starts multiple tasks in parallel. They got worse output by the third complex task because the human brain is depleted. Mandate parallel work, but only one complex task per engineer at a time.~355–385
14Team shape — two-to-four sub-streamsTwo-pizza teams generate 48 PRs; review becomes the bottleneck. Break into 2–4 person sub-streams with adoption-bound missions; cross-stream architects review key ADRs and integration.~385–435
15Does AI need good engineering practices?Made-up velocity-degradation graph. Causes: technical debt + success. Loose coupling, fast feedback (<15 min full test suite), and blast-radius–aware design matter more under AI velocity.~435–480
16Closing — find your principlesDecide what humans must understand vs. agents; decide vendor-lock tolerance; identify the real bottleneck idea→adoption; reject vanity metrics (tokens, commits, PRs); measure adoption. QR code for follow-up articles.~480–510

Terminology glossary (speaker's own framing)

  • Amplifier (AI as)"AI is an amplifier. So for the things that we did really well as a company, it allows us to do it much quicker. ... For the things that we did really badly ... then all it's done is amplified the things that we didn't do so well."
  • Architectural Decision Record (ADR)"about documenting your technical decisions along with your code so that you can someone can quickly learn why we made the decision so we don't keep on discussing them again." Under AI workflow: humans write them first, with diagrams; agents review subsequent PRs against them.
  • The producer black box — Three opaque layers: the harness/interface (Claude Code, OpenCode, Codex, agentic workflows), the model host, and the model itself. "Really your back box is made of three things that you don't understand."
  • Vanity metric"out of all that code ... it's all written by AI now. But how much of it was written without any human intervention?" Specifically called out as vanity: token usage, commits, PRs merged, % of code written by AI.
  • Blast radius"What is the worst thing that can go wrong in one instance of a pipeline after reduction happens" — design so one component going down still leaves most product functionality working.
  • You build it, you run it, you drive adoption — Batey's extension of the classic Werner Vogels DevOps mantra: "for now it's like you build it, you run it and you have to drive the adoption of that feature."
  • Sub-streams of two-to-four — Smaller-than-two-pizza units inside a 6–8-person team, each with an adoption-bound mission. "you get three pizzas because honestly you can't have a two tin subs you share pizzas."
  • Spec-driven frameworks — Tooling like Superpowers used during the planning/refinement phases. "superpowers, I would highly recommend it."

Named frameworks / concepts introduced

  1. The three pillars (from the abstract; the talk delivers on them in scrambled order):
    1. Path to Production at AI Speed — automated architectural checks, API/schema protection, static validation, testing for high change velocity.
    2. Training and Evaluating AI-Enabled Engineers — explicit coaching in AI workflows; rethinking what signals indicate effective AI-assisted development.
    3. Designing Workflow for Parallel Change — preview environments, isolated branches, architectural boundaries enabling concurrent change.
  2. Don't delegate systems thinking to agents — and the example list of system-level questions for a multi-tenant managed service ↔ on-prem agent connection.
  3. ADR-first workflow with AI — humans write ADRs first (with diagrams); agents enforce them on follow-up PRs.
  4. Producer black box (harness / host / model) — and the prescription to retain the ability to switch providers.
  5. Idea → Product Mgmt → Product Ownership → Build → Adoption — uneven AI acceleration across stages creates friction.
  6. Parallel work, one complex task at a time — mandate parallelism but cap cognitive load.
  7. Two-to-four-person sub-streams with adoption missions — replace 6–8-person two-pizza teams when AI inflates PR throughput.
  8. Closing principles checklist — what humans understand vs. agents; vendor-lock tolerance; identify real bottleneck; reject vanity metrics; define "good" concretely; hold AI accountable to it.

Open questions / not covered

  • Specific test taxonomy — Batey explicitly cut this: "I was going to go through an example of the types of tests we use but I'm not going to have time but feel taught in me after it if you want to." The talk does not enumerate the testing strategies despite the abstract promising them.
  • Concrete training-platform mechanics — he mentions the interactive training platform but doesn't detail curriculum or evaluation rubrics for AI-enabled engineers.
  • How "evaluating AI-enabled engineers" actually works — pillar #2 of the abstract is the thinnest in the delivered talk. Specific signals/metrics for effective AI-assisted development are not enumerated.
  • Hiring practices — abstract mentions hiring from scratch; the talk doesn't address interview design or hiring criteria.
  • Preview environments / isolated branches — promised in pillar #3 of the abstract but not walked through.
  • Specific tools beyond Superpowers — Claude Code, OpenCode, Codex are mentioned as examples of harnesses; no detailed comparison.
  • Concrete adoption metrics — Batey says "measure adoption" but doesn't define how.
  • Q&A — host says "I'm sorry there's no questions because I'm sure there are plenty for this. I'll stay here and I'll go outside if anyone's got any questions" — so no on-stage Q&A is captured.

talk-batey-building-product-teams-age-of-ai

README.md

tile.json